psi: stop relying on timer_pending for poll_work rescheduling
Subsystems (3)
Suren Baghdasaryan
maintainerPATCH v2 1/1 psi: stop relying on timer_pending for poll_work rescheduling
Psi polling mechanism is trying to minimize the number of wakeups to run psi_poll_work and is currently relying on timer_pending() to detect when this work is already scheduled. This provides a window of opportunity for psi_group_change to schedule an immediate psi_poll_work after poll_timer_fn got called but before psi_poll_work could reschedule itself. Below is the depiction of this entire window:
poll_timer_fn
wake_up_interruptible(&group->poll_wait);
psi_poll_worker
wait_event_interruptible(group->poll_wait, ...)
psi_poll_work
psi_schedule_poll_work
if (timer_pending(&group->poll_timer)) return;
...
mod_timer(&group->poll_timer, jiffies + delay);Prior to 461daba06bdc we used to rely on poll_scheduled atomic which was reset and set back inside psi_poll_work and therefore this race window was much smaller.
The larger window causes increased number of wakeups and our partners report visible power regression of ~10mA after applying 461daba06bdc. Bring back the poll_scheduled atomic and make this race window even narrower by resetting poll_scheduled only when we reach polling expiration time. This does not completely eliminate the possibility of extra wakeups caused by a race with psi_group_change however it will limit it to the worst case scenario of one extra wakeup per every tracking window (0.5s in the worst case).
By tracing the number of immediate rescheduling attempts performed by psi_group_change and the number of these attempts being blocked due to psi monitor being already active, we can assess the effects of this change:
Before the patch:
Run#1 Run#2 Run#3 Immediate reschedules attempted: 684365 1385156 1261240 Immediate reschedules blocked: 682846 1381654 1258682 Immediate reschedules (delta): 1519 3502 2558 Immediate reschedules (% of attempted): 0.22% 0.25% 0.20%
After the patch:
Run#1 Run#2 Run#3 Immediate reschedules attempted: 882244 770298 426218 Immediate reschedules blocked: 881996 769796 426074 Immediate reschedules (delta): 248 502 144 Immediate reschedules (% of attempted): 0.03% 0.07% 0.03%
The number of non-blocked immediate reschedules dropped from 0.22-0.25% to 0.03-0.07%. The drop is attributed to the decrease in the race window size and the fact that we allow this race only when psi monitors reach polling window expiration time.
---
changes in v2:
- Simplified atomic_cmpxchg() condition per PeterZ
- Added Tested-by:
include/linux/psi_types.h | 1 + kernel/sched/psi.c | 41 ++++++++++++++++++++++++++++----------- 2 files changed, 31 insertions(+), 11 deletions(-)
include/linux/psi_types.h +1 -0
@@ -158,6 +158,7 @@ struct psi_group {struct timer_list poll_timer;wait_queue_head_t poll_wait;atomic_t poll_wakeup; +atomic_t poll_scheduled;/* Protects data used by the monitor */ struct mutex trigger_lock;
kernel/sched/psi.c +30 -11
92 lines not on this page · raw message
Signature · 1 line
2.32.0.93.g670b81a890-goog
Quoted 52 lines · Suren Baghdasaryan
Psi polling mechanism is trying to minimize the number of wakeups to run
psi_poll_workand is currently relying ontimer_pending()to detect when this work is already scheduled. This provides a window of opportunity forpsi_group_changeto schedule an immediatepsi_poll_workafterpoll_timer_fngot called but beforepsi_poll_workcould reschedule itself. Below is the depiction of this entire window:
poll_timer_fn
wake_up_interruptible(&group->poll_wait);
psi_poll_worker
wait_event_interruptible(group->poll_wait, ...)
psi_poll_workpsi_schedule_poll_work if (timer_pending(&group->poll_timer)) return; ... mod_timer(&group->poll_timer, jiffies + delay);Prior to
461daba06bdcwe used to rely onpoll_scheduledatomic which was reset and set back insidepsi_poll_workand therefore this race window was much smaller.
The larger window causes increased number of wakeups and our partners report visible power regression of ~10mA after applying461daba06bdc. Bring back thepoll_scheduledatomic and make this race window even narrower by resettingpoll_scheduledonly when we reach polling expiration time. This does not completely eliminate the possibility of extra wakeups caused by a race withpsi_group_changehowever it will limit it to the worst case scenario of one extra wakeup per every tracking window (0.5s in the worst case).
By tracing the number of immediate rescheduling attempts performed bypsi_group_changeand the number of these attempts being blocked due to psi monitor being already active, we can assess the effects of this change:Before the patch:
Run#1 Run#2 Run#3 Immediate reschedules attempted: 684365 1385156 1261240 Immediate reschedules blocked: 682846 1381654 1258682 Immediate reschedules (delta): 1519 3502 2558 Immediate reschedules (% of attempted): 0.22% 0.25% 0.20%After the patch:
Run#1 Run#2 Run#3 Immediate reschedules attempted: 882244 770298 426218 Immediate reschedules blocked: 881996 769796 426074 Immediate reschedules (delta): 248 502 144 Immediate reschedules (% of attempted): 0.03% 0.07% 0.03%The number of non-blocked immediate reschedules dropped from 0.22-0.25% to 0.03-0.07%. The drop is attributed to the decrease in the race window size and the fact that we allow this race only when psi monitors reach polling window expiration time.
Quoted 6 lines
Fixes 461daba06bdc ("psi: eliminate kthread_worker from psi trigger scheduling mechanism")Reported-by Kathleen Chang <[email protected]>Reported-by Wenju Xu <[email protected]>Reported-by Jonathan Chen <[email protected]>Signed-off-by Suren Baghdasaryan <[email protected]>Tested-by SH Chen <[email protected]>
Peter Zijlstra
maintainerkernel/sched/psi.c static u64 update_triggers(struct psi_group *group, u64 now) · line 567
Quoted 3 lines · Suren Baghdasaryan
kernel/sched/psi.c +2
-0+/* cmpxchg should be called even when !force to set poll_scheduled */ +if (atomic_cmpxchg (&group->poll_scheduled,0 ,1 ) && !force)return ;
Why is that a cmpxchg() ?
Suren Baghdasaryan
maintainerQuoted 7 lines · Peter Zijlstra
kernel/sched/psi.c static u64 update_triggers(struct psi_group *group, u64 now) · line 567
Quoted 3 lines · Suren Baghdasaryan
kernel/sched/psi.c +2
-0+/* cmpxchg should be called even when !force to set poll_scheduled */ +if (atomic_cmpxchg (&group->poll_scheduled,0 ,1 ) && !force)return ;Why is that a
cmpxchg()?
We want to set poll_scheduled and proceed with rescheduling the timer unless it's already scheduled, so cmpxchg helps us to make that decision atomically. Or did I misunderstand your question?
Quoted 4 lines
--
To unsubscribe from this group and stop receiving emails from it, send an email to [email protected].
Peter Zijlstra
maintainerQuoted 7 lines · Suren Baghdasaryan
Quoted 6 lines · Peter Zijlstra
Quoted 3 lines · Suren Baghdasaryan
kernel/sched/psi.c +2
-0+/* cmpxchg should be called even when !force to set poll_scheduled */ +if (atomic_cmpxchg (&group->poll_scheduled,0 ,1 ) && !force)return ;
Quoted 5 lines
Quoted 1 line
Why is that a
cmpxchg()?We want to set
poll_scheduledand proceed with rescheduling the timer unless it's already scheduled, so cmpxchg helps us to make that decision atomically. Or did I misunderstand your question?
What's wrong with: atomic_xchg(&group->poll_scheduled, 1) ?
_______________________________________________
Linux-mediatek mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-mediatek
Suren Baghdasaryan
maintainerQuoted 9 lines · Peter Zijlstra
Quoted 7 lines · Suren Baghdasaryan
Quoted 6 lines · Peter Zijlstra
Quoted 3 lines · Suren Baghdasaryan
kernel/sched/psi.c +2
-0+/* cmpxchg should be called even when !force to set poll_scheduled */ +if (atomic_cmpxchg (&group->poll_scheduled,0 ,1 ) && !force)return ;
Quoted 7 lines
Quoted 5 lines
Quoted 1 line
Why is that a
cmpxchg()?We want to set
poll_scheduledand proceed with rescheduling the timer unless it's already scheduled, so cmpxchg helps us to make that decision atomically. Or did I misunderstand your question?What's wrong with:
atomic_xchg(&group->poll_scheduled, 1)?
Yes, since poll_scheduled can be only 0 or 1 atomic_xchg should work fine here. Functionally equivalent but I assume atomic_xchg() is more efficient due to no comparison.
Will respin v3 if the rest looks ok.
Quoted 4 lines
--
To unsubscribe from this group and stop receiving emails from it, send an email to [email protected].
Peter Zijlstra
maintainerQuoted 16 lines · Suren Baghdasaryan
Quoted 15 lines · Peter Zijlstra
Quoted 12 lines · Suren Baghdasaryan
Quoted 7 lines · Peter Zijlstra
kernel/sched/psi.c static u64 update_triggers(struct psi_group *group, u64 now) · line 567
Quoted 3 lines · Suren Baghdasaryan
kernel/sched/psi.c +2
-0+/* cmpxchg should be called even when !force to set poll_scheduled */ +if (atomic_cmpxchg (&group->poll_scheduled,0 ,1 ) && !force)return ;Why is that a
cmpxchg()?We want to set
poll_scheduledand proceed with rescheduling the timer unless it's already scheduled, so cmpxchg helps us to make that decision atomically. Or did I misunderstand your question?
Quoted 5 lines
Quoted 1 line
What's wrong with:
atomic_xchg(&group->poll_scheduled, 1)?Yes, since
poll_scheduledcan be only 0 or 1atomic_xchgshould work fine here. Functionally equivalent but I assumeatomic_xchg()is more efficient due to no comparison.
Mostly conceptually simpler; the cmpxchg-on-0 makes that you have to check if there's ever any state outside of {0,1}. The xchg() thing is the classical test-and-set pattern.
On top of all that, the cmpxchg() can fail, which brings ordering issues.
Typically, I think, you want to ensure that everything that happens before psi_schedule_poll_work() is visible to the work when it runs (also see Johannes' email). In case poll_scheduled is already 1, the cmpxchg will fail and *NOT* provide that ordering. Meaning the work might not observe the latest changes. xchg() doesn't have this subtlety.
Suren Baghdasaryan
maintainerQuoted 24 lines · Peter Zijlstra
Quoted 16 lines · Suren Baghdasaryan
Quoted 15 lines · Peter Zijlstra
Quoted 12 lines · Suren Baghdasaryan
Quoted 7 lines · Peter Zijlstra
kernel/sched/psi.c static u64 update_triggers(struct psi_group *group, u64 now) · line 567
Quoted 3 lines · Suren Baghdasaryan
kernel/sched/psi.c +2
-0+/* cmpxchg should be called even when !force to set poll_scheduled */ +if (atomic_cmpxchg (&group->poll_scheduled,0 ,1 ) && !force)return ;Why is that a
cmpxchg()?We want to set
poll_scheduledand proceed with rescheduling the timer unless it's already scheduled, so cmpxchg helps us to make that decision atomically. Or did I misunderstand your question?Quoted 5 lines
Quoted 1 line
What's wrong with:
atomic_xchg(&group->poll_scheduled, 1)?Yes, since
poll_scheduledcan be only 0 or 1atomic_xchgshould work fine here. Functionally equivalent but I assumeatomic_xchg()is more efficient due to no comparison.
Quoted 6 lines
Mostly conceptually simpler; the cmpxchg-on-0 makes that you have to check if there's ever any state outside of {0,1}. The
xchg()thing is the classical test-and-set pattern.On top of all that, the
cmpxchg()can fail, which brings ordering issues.
Oh, I see. That was my mistake. I was wrongly assuming that all RMW atomic operations are fully ordered but indeed, documentation states that:
``` - RMW operations that have a return value are fully ordered; - RMW operations that are conditional are unordered on FAILURE, otherwise the above rules apply. ```
So that's the actual functional difference here. Thanks for catching this and educating me!
Quoted 4 lines
Typically, I think, you want to ensure that everything that happens before
psi_schedule_poll_work()is visible to the work when it runs (also see Johannes' email).
Correct and I think I understand now the concern Johannes expressed.
Quoted 3 lines
In case
poll_scheduledis already 1, the
cmpxchg will fail and *NOT* provide that ordering. Meaning the work might not observe the latest changes.xchg()doesn't have this subtlety.
Got it.
So I think the modifications needed to this patch is:
1. replacing atomic_cmpxchg(&group->poll_scheduled, 0, 1) with atomic_chg(&group->poll_scheduled, 1)
2. an explicit smp_mb() barrier right after atomic_set(&group->poll_scheduled, 0) in psi_poll_work().
I think that should ensure the correct ordering here. If you folks agree I'll respin v3 with these changes (or maybe I should respin and we continue discussion with that version?).
Quoted 4 lines
--
To unsubscribe from this group and stop receiving emails from it, send an email to [email protected].
Suren Baghdasaryan
maintainerQuoted 62 lines · Suren Baghdasaryan
Quoted 24 lines · Peter Zijlstra
Quoted 16 lines · Suren Baghdasaryan
Quoted 15 lines · Peter Zijlstra
Quoted 12 lines · Suren Baghdasaryan
Quoted 7 lines · Peter Zijlstra
Review
Quoted 3 lines · Suren Baghdasaryan
(quoted diff) +2
-0+ /* cmpxchg should be called even when !force to set poll_scheduled */ + if (atomic_cmpxchg(&group->poll_scheduled, 0, 1) && !force) return;Why is that a
cmpxchg()?We want to set
poll_scheduledand proceed with rescheduling the timer unless it's already scheduled, so cmpxchg helps us to make that decision atomically. Or did I misunderstand your question?Quoted 5 lines
Quoted 1 line
What's wrong with:
atomic_xchg(&group->poll_scheduled, 1)?Yes, since
poll_scheduledcan be only 0 or 1atomic_xchgshould work fine here. Functionally equivalent but I assumeatomic_xchg()is more efficient due to no comparison.Quoted 6 lines
Mostly conceptually simpler; the cmpxchg-on-0 makes that you have to check if there's ever any state outside of {0,1}. The
xchg()thing is the classical test-and-set pattern.On top of all that, the
cmpxchg()can fail, which brings ordering issues.Oh, I see. That was my mistake. I was wrongly assuming that all RMW atomic operations are fully ordered but indeed, documentation states that:
``` - RMW operations that have a return value are fully ordered; - RMW operations that are conditional are unordered on FAILURE, otherwise the above rules apply. ```So that's the actual functional difference here. Thanks for catching this and educating me!
Quoted 4 lines
Typically, I think, you want to ensure that everything that happens before
psi_schedule_poll_work()is visible to the work when it runs (also see Johannes' email).Correct and I think I understand now the concern Johannes expressed.
Quoted 3 lines
In case
poll_scheduledis already 1, the
cmpxchg will fail and *NOT* provide that ordering. Meaning the work might not observe the latest changes.xchg()doesn't have this subtlety.Got it.
So I think the modifications needed to this patch is:
1. replacingatomic_cmpxchg(&group->poll_scheduled, 0, 1)withatomic_chg(&group->poll_scheduled, 1)
2. an explicitsmp_mb()barrier right afteratomic_set(&group->poll_scheduled, 0)inpsi_poll_work().
Quoted 3 lines
I think that should ensure the correct ordering here. If you folks agree I'll respin v3 with these changes (or maybe I should respin and we continue discussion with that version?).
To keep things moving I posted v3
(https://lore.kernel.org/patchwork/patch/1454547) with the changes I mentioned above. Let's keep discussing it there. Thanks!
Quoted 5 lines
Quoted 5 lines
--
To unsubscribe from this group and stop receiving emails from it, send an email to [email protected].
Johannes Weiner
maintainerQuoted 6 lines · Peter Zijlstra
kernel/sched/psi.c static u64 update_triggers(struct psi_group *group, u64 now) · line 567
Quoted 3 lines · Suren Baghdasaryan
kernel/sched/psi.c +2
-0+/* cmpxchg should be called even when !force to set poll_scheduled */ +if (atomic_cmpxchg (&group->poll_scheduled,0 ,1 ) && !force)return ;Why is that a
cmpxchg()?
I now realize you had already pointed that out, but I dismissed it in the context of poll_lock not being always taken after all.
But you're right, cmpxchg indeed seems inappropriate. xchg will do just fine for this binary toggle.
When it comes to ordering, looking at it again, I think we actually need ordering here that the seqcount doesn't provide. We have:
timer:
scheduled = 0smp_rmb()
x = state
scheduler:
state = ysmp_wmb()
if xchg(scheduled, 1) == 0
mod_timer()
Again, the requirement is that when the scheduler sees the timer as already or still pending, the timer must observe its state updates - otherwise we miss poll events.
The seqcount provides the wmb and rmb, but the scheduler-side read of @scheduled mustn't be reordered before the write to @state. Likewise, the timer-side read of @state also mustn't occur before the write to @scheduled.
AFAICS this is broken, not just in the patch, but also in the current code when timer_pending() on the scheduler side gets reordered. (Not sure if timer reading state can be reordered before the detach_timer() of its own expiration, but I don't see full ordering between them.)
So it seems to me we need the ordered atomic_xchg() on the scheduler side, and on the timer side an smp_mb() after we set scheduled to 0.
Suren Baghdasaryan
maintainerQuoted 38 lines · Johannes Weiner
Quoted 6 lines · Peter Zijlstra
kernel/sched/psi.c static u64 update_triggers(struct psi_group *group, u64 now) · line 567
Quoted 3 lines · Suren Baghdasaryan
kernel/sched/psi.c +2
-0+/* cmpxchg should be called even when !force to set poll_scheduled */ +if (atomic_cmpxchg (&group->poll_scheduled,0 ,1 ) && !force)return ;Why is that a
cmpxchg()?I now realize you had already pointed that out, but I dismissed it in the context of
poll_locknot being always taken after all.But you're right, cmpxchg indeed seems inappropriate. xchg will do just fine for this binary toggle.
When it comes to ordering, looking at it again, I think we actually need ordering here that the seqcount doesn't provide. We have:
timer:
scheduled = 0smp_rmb()
x = statescheduler:
state = ysmp_wmb()
ifxchg(scheduled, 1)== 0
mod_timer()Again, the requirement is that when the scheduler sees the timer as already or still pending, the timer must observe its state updates - otherwise we miss poll events.
The seqcount provides the wmb and rmb, but the scheduler-side read of @scheduled mustn't be reordered before the write to @state. Likewise, the timer-side read of @state also mustn't occur before the write to @scheduled.
Quoted 7 lines
AFAICS this is broken, not just in the patch, but also in the current code when
timer_pending()on the scheduler side gets reordered. (Not sure if timer reading state can be reordered before thedetach_timer()of its own expiration, but I don't see full ordering between them.)So it seems to me we need the ordered
atomic_xchg()on the scheduler side, and on the timer side ansmp_mb()after we set scheduled to 0.
Thanks for the analysis Johannes. Let me dwell on it a bit.